Filter std:: instantiations out of Halide.dll's export list

CMakeLists.txt

@@ -95,7 +95,8 @@ if (CMAKE_SYSTEM_NAME MATCHES "Windows")
     endif ()
 endif ()
 
-# Export all symbols on Windows to match GCC/Clang behavior on Linux/macOS
+# Export all symbols on Windows to match GCC/Clang behavior on Linux/macOS.
+# (With MSVC, src/CMakeLists.txt filters Halide.dll's list to stay under the linker's limit.)
 set(CMAKE_WINDOWS_EXPORT_ALL_SYMBOLS ON)
 
 # Require standard C++17

cmake/FilterDllExports.cmake

@@ -0,0 +1,173 @@
+# Filters a module-definition (.def) export list produced by
+# `cmake -E __create_def`, dropping symbols that are not ours to export.
+#
+# Usage: cmake -DIN=<in.def> -DOUT=<out.def> -P FilterDllExports.cmake
+#
+# CMAKE_WINDOWS_EXPORT_ALL_SYMBOLS exports every external symbol in Halide's
+# object files, including COMDAT instantiations of std:: templates. The list is
+# close to link.exe's limit of 65535 exports (LNK1189), so we drop symbols whose
+# *enclosing scope* (the outermost namespace/class in the MSVC-decorated name)
+# is std, stdext, Concurrency, llvm or clang. Clients instantiate their own
+# copies of those templates. A substring match on "@std@@" would be wrong, as
+# it also appears in the parameter encoding of Halide functions taking, e.g.,
+# std::string.
+#
+# Whenever a name cannot be confidently reduced, it is kept.
+#
+# Performance: this runs on every link over ~65k lines, and character-level
+# loops in CMake are far too slow. Instead we rely on the fact that a template
+# argument list containing no nested template is a regular language. Innermost
+# templates are repeatedly collapsed into a plain fragment until the qualified
+# name is flat, then the outermost scope is read with a single regex.
+
+if (NOT DEFINED IN OR NOT DEFINED OUT)
+    message(FATAL_ERROR "Usage: cmake -DIN=<in.def> -DOUT=<out.def> -P FilterDllExports.cmake")
+endif ()
+
+set(EXPORT_LIMIT 65535)
+set(MAX_PASSES 8)
+
+# -- Building blocks of the mangled-name grammar --------------------------------
+
+# An identifier fragment, e.g. `Halide@`.
+set(ident [=[[A-Za-z_][A-Za-z0-9_]*@]=])
+# A qualified name: fragments (identifiers or one-digit back-references)
+# terminated by `@`, e.g. `Func@Halide@@`. Used after V/U/T/W4 in a type.
+set(qual "((${ident}|[0-9])*@)")
+# One template argument. Deliberately excludes anything that contains `@` or
+# `?` in a way we do not model (function types `6`/`8`, arrays `Y`, `?`...), so
+# such templates fail to match and their symbols are kept.
+string(CONCAT arg
+    "(" "[VUT]${qual}"  # class/struct/union by name
+    "|" "W[0-9]${qual}"  # enum
+    "|" [=[\$\$[A-Za-z]]=]  # $$Q, $$R, $$T, $$C...
+    "|" [=[\$0[0-9A-P]*@]=]  # non-negative integer value
+    "|" "_[A-Za-z0-9]"  # _N, _J, ...
+    "|" "[0-579A-SXZ]"  # builtins, modifiers, back-references
+    ")"
+)
+
+# An innermost template-id `?$name@<args>@`; \1 is the name.
+set(template_re [=[\?\$([A-Za-z_][A-Za-z0-9_]*)@]=])
+string(APPEND template_re "${arg}*@")
+
+# The outermost scope of a flat qualified name: at least one fragment, then a
+# final identifier, terminated by `@@`. \3 is the outermost scope.
+set(scope_re "^(${ident}|[0-9])+([A-Za-z_][A-Za-z0-9_]*)@@")
+# A qualified name of a single fragment, i.e. an unscoped global.
+set(global_re "^[A-Za-z_][A-Za-z0-9_]*@@")
+
+set(dropped_scope_re "^(std|stdext|Concurrency|llvm|clang)$")
+
+# -- Main loop ------------------------------------------------------------------
+
+file(READ "${IN}" content)
+string(REPLACE "\r" "" content "${content}")
+string(REPLACE "\n" ";" lines "${content}")
+
+set(out "")
+set(total 0)
+set(dropped 0)
+set(kept 0)
+set(unparsed 0)
+
+foreach (line IN LISTS lines)
+    if (NOT line MATCHES "[^ \t]")
+        continue()
+    endif ()
+    if (line MATCHES "^[ \t]*(EXPORTS|LIBRARY)")
+        string(APPEND out "${line}\n")
+        continue()
+    endif ()
+
+    string(REGEX REPLACE "^[ \t]+" "" symbol "${line}")
+    string(REGEX REPLACE "[ \t].*$" "" symbol "${symbol}")
+    math(EXPR total "${total} + 1")
+
+    set(decision KEEP)
+
+    # Strip the special-name prefix, leaving the qualified name. Anything
+    # undecorated (extern "C") is kept as is.
+    set(body "")
+    if (symbol MATCHES "^\\?")
+        if (symbol MATCHES "^\\?\\?_C")
+            # String literal.
+        elseif (symbol MATCHES "^\\?\\?\\$")
+            # Template function: the name is `?$name@...`
+            string(SUBSTRING "${symbol}" 1 -1 body)
+        elseif (symbol MATCHES "^\\?\\?_R0\\?A[VU]")
+            string(SUBSTRING "${symbol}" 8 -1 body)
+        elseif (symbol MATCHES "^\\?\\?_R4")
+            string(SUBSTRING "${symbol}" 5 -1 body)
+        elseif (symbol MATCHES "^\\?\\?_R")
+            set(decision UNPARSED)
+        elseif (symbol MATCHES "^\\?\\?__[0-9A-Za-z]")
+            string(SUBSTRING "${symbol}" 5 -1 body)
+        elseif (symbol MATCHES "^\\?\\?_[0-9A-Za-z]")
+            string(SUBSTRING "${symbol}" 4 -1 body)
+        elseif (symbol MATCHES "^\\?\\?[0-9A-Z]")
+            string(SUBSTRING "${symbol}" 3 -1 body)
+        elseif (symbol MATCHES "^\\?\\?")
+            set(decision UNPARSED)
+        else ()
+            string(SUBSTRING "${symbol}" 1 -1 body)
+        endif ()
+    endif ()
+
+    if (NOT "${body}" STREQUAL "")
+        foreach (pass RANGE 1 ${MAX_PASSES})
+            if (NOT body MATCHES "\\?\\$")
+                break()
+            endif ()
+            set(before "${body}")
+            string(REGEX REPLACE "${template_re}" "\\1@" body "${body}")
+            if (body STREQUAL before)
+                break()
+            endif ()
+        endforeach ()
+
+        if (body MATCHES "${scope_re}")
+            if ("${CMAKE_MATCH_2}" MATCHES "${dropped_scope_re}")
+                set(decision DROP)
+            endif ()
+        elseif (NOT body MATCHES "${global_re}")
+            set(decision UNPARSED)
+        endif ()
+    endif ()
+
+    if (decision STREQUAL DROP)
+        math(EXPR dropped "${dropped} + 1")
+        continue()
+    elseif (decision STREQUAL UNPARSED)
+        math(EXPR unparsed "${unparsed} + 1")
+    endif ()
+    math(EXPR kept "${kept} + 1")
+    string(APPEND out "${line}\n")
+endforeach ()
+
+if (total EQUAL 0)
+    message(FATAL_ERROR "No exports found in ${IN}")
+endif ()
+
+file(WRITE "${OUT}" "${out}")
+
+message(
+    VERBOSE "Halide.dll exports: ${total} total, ${dropped} filtered out, ${kept} remaining "
+    "(${unparsed} kept only because their names could not be parsed); limit is ${EXPORT_LIMIT}"
+)
+
+if (kept GREATER EXPORT_LIMIT)
+    message(
+        FATAL_ERROR "Halide.dll would export ${kept} symbols after filtering, exceeding the "
+        "linker limit of ${EXPORT_LIMIT} (LNK1189). Reduce the number of exported "
+        "symbols or extend cmake/FilterDllExports.cmake."
+    )
+endif ()
+
+math(EXPR warn_threshold "${EXPORT_LIMIT} * 95 / 100")
+if (kept GREATER warn_threshold)
+    message(
+        WARNING "Halide.dll exports ${kept} symbols after filtering, over 95% of the linker "
+        "limit of ${EXPORT_LIMIT}."
+    )
+endif ()

src/CMakeLists.txt

@@ -628,6 +628,50 @@ target_export_script(
     GNU_LD "${CMAKE_CURRENT_LIST_DIR}/exported_symbols.ldscript"
 )
 
+##
+# Trim Halide.dll's export list on MSVC
+##
+
+# CMAKE_WINDOWS_EXPORT_ALL_SYMBOLS exports every external symbol in our object
+# files, including COMDAT instantiations of std:: templates. That list is close
+# to link.exe's limit of 65535 exports (LNK1189), so instead we generate the
+# same list ourselves, drop symbols whose enclosing scope isn't ours (see
+# cmake/FilterDllExports.cmake), and hand the result to the linker.
+#
+# `cmake -E __create_def` is an internal CMake command, but it is exactly what
+# CMake itself runs for WINDOWS_EXPORT_ALL_SYMBOLS.
+if (WIN32 AND BUILD_SHARED_LIBS AND CMAKE_CXX_COMPILER_ID STREQUAL "MSVC")
+    set(Halide_EXPORTS_DIR "${CMAKE_CURRENT_BINARY_DIR}/halide_exports/$<CONFIG>")
+
+    set_target_properties(Halide PROPERTIES WINDOWS_EXPORT_ALL_SYMBOLS OFF)
+
+    file(GENERATE
+        OUTPUT "${Halide_EXPORTS_DIR}/objects.txt"
+        CONTENT "$<JOIN:$<TARGET_OBJECTS:Halide>;$<TARGET_OBJECTS:Halide_initmod>,\n>\n"
+    )
+
+    # Export counts are reported at --log-level=VERBOSE; forward the setting.
+    set(Halide_EXPORTS_LOG_LEVEL "")
+    if (CMAKE_MESSAGE_LOG_LEVEL)
+        set(Halide_EXPORTS_LOG_LEVEL "--log-level=${CMAKE_MESSAGE_LOG_LEVEL}")
+    endif ()
+
+    add_custom_command(
+        TARGET Halide
+        PRE_LINK
+        COMMAND
+            "${CMAKE_COMMAND}" -E __create_def "${Halide_EXPORTS_DIR}/full.def"
+            "${Halide_EXPORTS_DIR}/objects.txt"
+        COMMAND
+            "${CMAKE_COMMAND}" ${Halide_EXPORTS_LOG_LEVEL} "-DIN=${Halide_EXPORTS_DIR}/full.def"
+            "-DOUT=${Halide_EXPORTS_DIR}/filtered.def" -P
+            "${Halide_SOURCE_DIR}/cmake/FilterDllExports.cmake"
+        COMMENT "Generating filtered export list for Halide.dll"
+        VERBATIM
+    )
+    target_link_options(Halide PRIVATE "/DEF:${Halide_EXPORTS_DIR}/filtered.def")
+endif ()
+
 ##
 # Set compiler options for libHalide
 ##

test/CMakeLists.txt

@@ -84,3 +84,12 @@ Halide_feature(WITH_TEST_DEBUGGER "Build debugger-integration tests" ON)
 if (WITH_TEST_DEBUGGER)
     add_subdirectory(debugger)
 endif ()
+
+add_test(
+    NAME filter_dll_exports
+    COMMAND
+        "${CMAKE_COMMAND}" "-DSCRIPT=${Halide_SOURCE_DIR}/cmake/FilterDllExports.cmake"
+        "-DDIR=${CMAKE_CURRENT_SOURCE_DIR}/filter_dll_exports"
+        "-DOUT=${CMAKE_CURRENT_BINARY_DIR}/filter_dll_exports.out.def" -P
+        "${CMAKE_CURRENT_SOURCE_DIR}/filter_dll_exports/run.cmake"
+)

test/filter_dll_exports/expected.def

@@ -0,0 +1,27 @@
+EXPORTS
+	?compile_to_bitcode@Func@Halide@@QAEXABV?$basic_string@DU?$char_traits@D@std@@V?$allocator@D@2@@std@@ABV?$vector@VArgument@Halide@@V?$allocator@VArgument@Halide@@@std@@@4@0ABVTarget@2@@Z
+	?name@Func@Halide@@QBEABV?$basic_string@DU?$char_traits@D@std@@V?$allocator@D@2@@std@@XZ
+	??0Func@Halide@@QAE@ABV?$basic_string@DU?$char_traits@D@std@@V?$allocator@D@2@@std@@@Z
+	?const_true@Halide@@YA?AVExpr@1@H@Z
+	??$make_const@H@Internal@Halide@@YA?AVExpr@1@VType@1@H@Z
+	??0?$IntrusivePtr@VBaseExprNode@Internal@Halide@@@Internal@Halide@@QAE@XZ
+	??1?$IntrusivePtr@VBaseExprNode@Internal@Halide@@@Internal@Halide@@QAE@XZ
+	??_7?$_Func_impl_no_alloc@V<lambda_1>@?0??foo@Halide@@YAXXZ@XH@std@@6B@
+	??_G?$_Func_impl_no_alloc@P6AXH@ZXH@std@@UAEPAXI@Z
+	??_7Error@Halide@@6B@
+	??_R0?AVError@Halide@@@8
+	?global_table@Internal@Halide@@3PAHA	 DATA
+	?instances@?$Registry@VFoo@Halide@@@Internal@Halide@@2V?$vector@PAVFoo@Halide@@V?$allocator@PAVFoo@Halide@@@std@@@std@@A	 DATA
+	?free_helper@?A0x1234abcd@@YAXXZ
+	??$foo@V<lambda_1>@?0??bar@Halide@@YAXXZ@@std@@YAXXZ
+	?vec_size@Halide@@YAIABV?$vector@HV?$allocator@H@std@@@std@@@Z
+	?callback@Halide@@YAXP6AXABV?$vector@HV?$allocator@H@std@@@std@@@Z@Z
+	??$apply@P6AXH@Z@Halide@@YAXXZ
+	??$bar@$00@@Halide@@YAXXZ
+	??$baz@U?$array@H$01@@std@@@Halide@@YAXXZ
+	halide_malloc
+	halide_error
+	_halide_buffer_get_host
+	halide_pipeline_finalize	 DATA
+	main
+	?f@@YAXXZ

test/filter_dll_exports/input.def

@@ -0,0 +1,42 @@
+EXPORTS
+	?compile_to_bitcode@Func@Halide@@QAEXABV?$basic_string@DU?$char_traits@D@std@@V?$allocator@D@2@@std@@ABV?$vector@VArgument@Halide@@V?$allocator@VArgument@Halide@@@std@@@4@0ABVTarget@2@@Z
+	?name@Func@Halide@@QBEABV?$basic_string@DU?$char_traits@D@std@@V?$allocator@D@2@@std@@XZ
+	??0Func@Halide@@QAE@ABV?$basic_string@DU?$char_traits@D@std@@V?$allocator@D@2@@std@@@Z
+	?const_true@Halide@@YA?AVExpr@1@H@Z
+	??$make_const@H@Internal@Halide@@YA?AVExpr@1@VType@1@H@Z
+	??0?$IntrusivePtr@VBaseExprNode@Internal@Halide@@@Internal@Halide@@QAE@XZ
+	??1?$IntrusivePtr@VBaseExprNode@Internal@Halide@@@Internal@Halide@@QAE@XZ
+	?_Tidy@?$vector@VFunc@Halide@@V?$allocator@VFunc@Halide@@@std@@@std@@AAEXXZ
+	??0?$vector@VFunc@Halide@@V?$allocator@VFunc@Halide@@@std@@@std@@QAE@XZ
+	??1?$vector@VFunc@Halide@@V?$allocator@VFunc@Halide@@@std@@@std@@QAE@XZ
+	??$_Emplace_reallocate@ABVFunc@Halide@@@?$vector@VFunc@Halide@@V?$allocator@VFunc@Halide@@@std@@@std@@QAEPAVFunc@Halide@@QAV23@ABV23@@Z
+	??$_Uninitialized_move@PAVFunc@Halide@@PAV12@V?$allocator@VFunc@Halide@@@std@@@std@@YAPAVFunc@Halide@@QAV12@0PAV12@AAV?$allocator@VFunc@Halide@@@0@@Z
+	??_7?$_Func_impl_no_alloc@V<lambda_1>@?0??foo@Halide@@YAXXZ@XH@std@@6B@
+	??_G?$_Func_impl_no_alloc@P6AXH@ZXH@std@@UAEPAXI@Z
+	??1?$map@HHU?$less@H@std@@V?$allocator@U?$pair@$$CBHH@std@@@2@@std@@QAE@XZ
+	??_7?$_Ref_count_obj2@VFoo@Halide@@@std@@6B@
+	??_R0?AV?$_Ref_count_obj2@VFoo@Halide@@@std@@@8
+	??_R4?$_Ref_count_obj2@VFoo@Halide@@@std@@6B@
+	??_7Error@Halide@@6B@
+	??_R0?AVError@Halide@@@8
+	??$_Get_size_of_n@$00@@std@@YAII@Z
+	?_Id@?$numpunct@D@std@@2V0locale@2@A	 DATA
+	?global_table@Internal@Halide@@3PAHA	 DATA
+	?instances@?$Registry@VFoo@Halide@@@Internal@Halide@@2V?$vector@PAVFoo@Halide@@V?$allocator@PAVFoo@Halide@@@std@@@std@@A	 DATA
+	?_Xlength_error@std@@YAXPBD@Z
+	?free_helper@?A0x1234abcd@@YAXXZ
+	??$foo@V<lambda_1>@?0??bar@Halide@@YAXXZ@@std@@YAXXZ
+	?vec_size@Halide@@YAIABV?$vector@HV?$allocator@H@std@@@std@@@Z
+	?callback@Halide@@YAXP6AXABV?$vector@HV?$allocator@H@std@@@std@@@Z@Z
+	??$apply@P6AXH@Z@Halide@@YAXXZ
+	??$bar@$00@@Halide@@YAXXZ
+	??$baz@U?$array@H$01@@std@@@Halide@@YAXXZ
+	??$sz@$$T@std@@YAXXZ
+	??$sz@_N@std@@YAXXZ
+	??$sz@W4Kind@Halide@@@std@@YAXXZ
+	halide_malloc
+	halide_error
+	_halide_buffer_get_host
+	halide_pipeline_finalize	 DATA
+	main
+	?f@@YAXXZ

test/filter_dll_exports/run.cmake

@@ -0,0 +1,17 @@
+# Runs cmake/FilterDllExports.cmake on input.def and compares against expected.def.
+# Usage: cmake -DSCRIPT=<FilterDllExports.cmake> -DDIR=<this dir> -DOUT=<scratch file> -P run.cmake
+execute_process(
+    COMMAND
+        "${CMAKE_COMMAND}" --log-level=VERBOSE "-DIN=${DIR}/input.def" "-DOUT=${OUT}" -P "${SCRIPT}"
+    RESULT_VARIABLE rc
+)
+if (NOT rc EQUAL 0)
+    message(FATAL_ERROR "FilterDllExports.cmake failed: ${rc}")
+endif ()
+execute_process(
+    COMMAND "${CMAKE_COMMAND}" -E compare_files "${DIR}/expected.def" "${OUT}"
+    RESULT_VARIABLE rc
+)
+if (NOT rc EQUAL 0)
+    message(FATAL_ERROR "FilterDllExports.cmake output differs from expected.def")
+endif ()